[CSSPGO][llvm-profgen] Trim and merge context beforehand to reduce memory usage

[wlei]

Summary

Currently we use a centralized string map(StringMap<FunctionSamples> ProfileMap) to store the profile while populating the sample, which might cause the memory usage bottleneck. I saw in an extreme case, there are thousands of samples whose context stack depth is >= 100. The memory consumption can be greater than 100GB.

As here the context is used for inlining, we can assume we won't have so many of inlinees keeping inlined at the same root function, so this change tried to cap the context stack and merge the samples for peak memory reduction and this is done after recursion compression.

The default value is -1 meaning no depth limit, in the future we can tune to a smaller one.

[wlei]

add to algin with line number based context

[wenlei]

llvm/tools/llvm-profgen/ProfileGenerator.cpp

53

I think we could unify the switch names, e.g. csprof-max-context-depth and csprof-max-cold-context-depth?

619

Since getExpandedContextStr only covers line-based profile, for probe we rely on the trimming here in profile generation, which is later then where we do the trimming for line-based profile. Do we see peak memory drop if we trim the context in profile generation instead of during unwinder?

llvm/tools/llvm-profgen/ProfileGenerator.h

73

nit: bottom-up order in stack is usually callers-callee order, from bottom can be confusing as it means we trim callees which is not the case.

also suggest rename capContextStack to trimContext.

[hoy]

llvm/tools/llvm-profgen/PerfReader.cpp

23

Nit: move this into ProfileGenerator.h to reduct the number of declarations?

llvm/tools/llvm-profgen/ProfileGenerator.cpp

53

Thanks for working on this. We probably do not inline so many levels of functions. But would be good to run through some perf testing or to turn this off by default.

[wlei]

addressing Wenlei and Hongtao's feedback

[wlei]

llvm/tools/llvm-profgen/PerfReader.cpp

23

fixed!

llvm/tools/llvm-profgen/ProfileGenerator.cpp

53

Sounds good, will collect the statistic of the max inline depth in SampleProfile inliner on some benchmarks and change to that one, maybe 10 is good enough.

619

Here it did for both, one during unwinder(see the one in PerfReader.cpp) and one here.

The answer is yes, it's better than unwinder only, here are some data:

10 depth for both: 17GB
10 depth for unwinder only: 26GB
20 depth for both: 42GB
20 depth for unwinder only: 49GB

llvm/tools/llvm-profgen/ProfileGenerator.h

73

Fixed!

[wlei]

fix lint

[wlei]

llvm/tools/llvm-profgen/ProfileGenerator.cpp

53

Here is the max inline depth(a inline b, then b inline c, the depth is 2) in SPEC2017 monoLTO pass2 (turn on all inliners).

508.namd_r  5
510.parest_r 21
511.povray_r  8
526.blender_r 15
600.perlbench_s 8
602.gcc_s 21 
605.mcf_s 5
620.omnetpp_s 18
623.xalancbmk_s 26
625.x264_s 7
631.deepsjeng_s 5
638.imagick_s 10
641.leela_s 16
644.nab_s 5
657.xz_s 7

and for the clang-10 pass1 binary(I don't have pass2 binary), the max inline depth is 51!

it's really more inlining than I thought. so I agree with you to turn it off(-1) by default.

[wlei]

default the value to -1

[wenlei]

llvm/tools/llvm-profgen/ProfileGenerator.cpp

53

Sounds reasonable. If we run into such situation more often, we could also try to have another level of aggregation by leaf frame from stack sample, then we can tell some contexts are cold before unwinding, and dynamically trim those cold context during unwinding.

Can we make the description and variable name consistent with CSProfColdContextFrameDepth too?

[wenlei]

llvm/tools/llvm-profgen/ProfileGenerator.cpp

53

to be specific: CSProfMaxContextDepth, CSProfMaxColdContextDepth. "Keep the last K frames while merging [cold] profile ..." otherwise the change looks good.

[spupyrev]

llvm/tools/llvm-profgen/ProfileGenerator.cpp

74

The default value of "no trimming" sounds strange to me. Do we know if trimming may affect performance? If yes, we shouldn't trim. If no, we should have a reasonable bound here.

[wlei]

rewrite the switch's variable and description

[wlei]

llvm/tools/llvm-profgen/ProfileGenerator.cpp

74

Do we know if trimming may affect performance?

It depends on how compiler consumes the context, if the profile context is [a @ b @ c @ d] but compiler only inline d to c and stop inlining at c, then the [a @ b @] can be trimmed. Another things is the profile size tradeoff, before we did merge and trim cold profile after all samples is generated. This patch tried to move the merging ahead as we encountered the memory issues. it doesn't drop the sample.

If no, we should have a reasonable bound here.

Good question, but I don't know what's the reasonable bound yet. As you see the above comments, SPEC benchmark's max inline depth can vary from 5 to 26, the clang-10 binary even have a 51 max depth. and we didn't get memory issue for SPEC, so SPEC can set to -1.

so it's the tradeoff between many things, what's in my mind is to default it -1 and we can tune it on-demand. what do you think?

[hoy]

llvm/tools/llvm-profgen/ProfileGenerator.cpp

74

Default -1 (means no trimming) sounds good to me. An otherwise decent default value would need quite some experiments to justify.

[spupyrev]

llvm/tools/llvm-profgen/ProfileGenerator.cpp

74

I dislike tunable flags that may or may not affect the performance. I assume most (all?) of the end-users won't know about the flag and will utilize the default value. So whatever you set here will be used in the vast majority of cases.

If course, if you do plan to run some perf experiments in the (near) future and tune the default value, then the current value of "-1" is OK.

[wlei]

llvm/tools/llvm-profgen/ProfileGenerator.cpp

74

I dislike tunable flags that may or may not affect the performance. I assume most (all?) of the end-users won't know about the flag and will utilize the default value. So whatever you set here will be used in the vast majority of cases.

That makes sense, thanks! Will tune it, maybe clang binary is the good place as it has the deep call stack.

llvm/test/tools/llvm-profgen/merge-cold-profile.test

 ; Used the data from recursion-compression.test, refer it for the unmerged output
 ; RUN: llvm-profgen --format=text --perfscript=%S/Inputs/recursion-compression-pseudoprobe.perfscript --binary=%S/Inputs/recursion-compression-pseudoprobe.perfbin --output=%t1 --compress-recursion=-1 --profile-summary-cold-count=8
 ; RUN: FileCheck %s --input-file %t1
 
 ; Test --csprof-trim-cold-context=0
 ; RUN: llvm-profgen --format=text --perfscript=%S/Inputs/recursion-compression-pseudoprobe.perfscript --binary=%S/Inputs/recursion-compression-pseudoprobe.perfbin --output=%t2 --compress-recursion=-1 --profile-summary-cold-count=100 --csprof-trim-cold-context=0
 ; RUN: FileCheck %s --input-file %t2 --check-prefix=CHECK-KEEP-COLD
 
 ; Test --csprof-merge-cold-context=0
 ; RUN: llvm-profgen --format=text --perfscript=%S/Inputs/recursion-compression-pseudoprobe.perfscript --binary=%S/Inputs/recursion-compression-pseudoprobe.perfbin --output=%t3 --compress-recursion=-1 --profile-summary-cold-count=10 --csprof-merge-cold-context=0
 ; RUN: FileCheck %s --input-file %t3 --check-prefix=CHECK-UNMERGED
 
 ; Test --csprof-frame-depth-for-cold-context
-; RUN: llvm-profgen --format=text --perfscript=%S/Inputs/recursion-compression-pseudoprobe.perfscript --binary=%S/Inputs/recursion-compression-pseudoprobe.perfbin --output=%t2 --compress-recursion=-1 --profile-summary-cold-count=100 --csprof-trim-cold-context=0 --csprof-frame-depth-for-cold-context=2
+; RUN: llvm-profgen --format=text --perfscript=%S/Inputs/recursion-compression-pseudoprobe.perfscript --binary=%S/Inputs/recursion-compression-pseudoprobe.perfbin --output=%t2 --compress-recursion=-1 --profile-summary-cold-count=100 --csprof-trim-cold-context=0 --csprof-max-cold-context-depth=2
 ; RUN: FileCheck %s --input-file %t2 --check-prefix=CHECK-COLD-CONTEXT-LENGTH
 
 ; CHECK:     [fa]:14:4
 ; CHECK-NEXT: 1: 4
 ; CHECK-NEXT: 3: 4
 ; CHECK-NEXT: 4: 2
 ; CHECK-NEXT: 5: 1
 ; CHECK-NEXT: 6: 0

llvm/test/tools/llvm-profgen/recursion-compression-noprobe.test

 ; Firstly test uncompression(--compress-recursion=0)
 ; RUN: llvm-profgen --format=text --perfscript=%S/Inputs/recursion-compression-noprobe.perfscript --binary=%S/Inputs/recursion-compression-noprobe.perfbin --output=%t --compress-recursion=0 --profile-summary-cold-count=0
 ; RUN: FileCheck %s --input-file %t -check-prefix=CHECK-UNCOMPRESS
 ; RUN: llvm-profgen --format=text --perfscript=%S/Inputs/recursion-compression-noprobe.perfscript --binary=%S/Inputs/recursion-compression-noprobe.perfbin --output=%t --profile-summary-cold-count=0
 ; RUN: FileCheck %s --input-file %t
+; RUN: llvm-profgen --format=text --perfscript=%S/Inputs/recursion-compression-noprobe.perfscript --binary=%S/Inputs/recursion-compression-noprobe.perfbin --output=%t --compress-recursion=0 --profile-summary-cold-count=0 --csprof-max-context-depth=2
+; RUN: FileCheck %s --input-file %t -check-prefix=CHECK-MAX-CTX-DEPTH
 
 ; CHECK-UNCOMPRESS:[main:1 @ foo:3 @ fa:2 @ fb]:48:0
 ; CHECK-UNCOMPRESS: 1: 11
 ; CHECK-UNCOMPRESS: 2: 1 fa:1
 ; CHECK-UNCOMPRESS:[main:1 @ foo:3 @ fa]:24:0
 ; CHECK-UNCOMPRESS: 1: 1
 ; CHECK-UNCOMPRESS: 2: 13 fb:11
 ; CHECK-UNCOMPRESS:[main:1 @ foo:3 @ fa:2 @ fb:2 @ fa]:7:0
 ; CHECK-UNCOMPRESS: 1: 1
 ; CHECK-UNCOMPRESS: 2: 2 fb:1
 ; CHECK-UNCOMPRESS:[main:1 @ foo]:7:0
 ; CHECK-UNCOMPRESS: 2: 1
 ; CHECK-UNCOMPRESS: 3: 2 fa:1
 ; CHECK-UNCOMPRESS:[main:1 @ foo:3 @ fa:2 @ fb:2 @ fa:2 @ fb:2 @ fa]:2:0
 ; CHECK-UNCOMPRESS: 4: 1
 ; CHECK-UNCOMPRESS:[main:1 @ foo:3 @ fa:2 @ fb:2 @ fa:2 @ fb]:2:0
 ; CHECK-UNCOMPRESS: 2: 1 fa:1
 
+; CHECK-MAX-CTX-DEPTH:[foo:3 @ fa:2 @ fb]:47:0
+; CHECK-MAX-CTX-DEPTH: 1: 11
+; CHECK-MAX-CTX-DEPTH:[main:1 @ foo:3 @ fa]:13:0
+; CHECK-MAX-CTX-DEPTH: 1: 1
+; CHECK-MAX-CTX-DEPTH: 2: 2
+; CHECK-MAX-CTX-DEPTH:[fa:2 @ fb:2 @ fa]:8:0
+; CHECK-MAX-CTX-DEPTH: 1: 1
+; CHECK-MAX-CTX-DEPTH: 2: 1
+; CHECK-MAX-CTX-DEPTH: 4: 1
+; CHECK-MAX-CTX-DEPTH:[main:1 @ foo]:7:0
+; CHECK-MAX-CTX-DEPTH: 2: 1
+; CHECK-MAX-CTX-DEPTH: 3: 2 fa:1
+; CHECK-MAX-CTX-DEPTH:[fb:2 @ fa:2 @ fb]:1:0
+
 
 ; CHECK: [main:1 @ foo:3 @ fa:2 @ fb]:48:0
 ; CHECK:  1: 11
 ; CHECK:  2: 1 fa:1
 ; CHECK: [main:1 @ foo:3 @ fa]:24:0
 ; CHECK:  1: 1
 ; CHECK:  2: 13 fb:11
 ; CHECK: [main:1 @ foo:3 @ fa:2 @ fb:2 @ fa]:9:0

llvm/test/tools/llvm-profgen/recursion-compression-pseudoprobe.test

 ; Firstly test uncompression(--compress-recursion=0)
 ; RUN: llvm-profgen --format=text --perfscript=%S/Inputs/recursion-compression-pseudoprobe.perfscript --binary=%S/Inputs/recursion-compression-pseudoprobe.perfbin --output=%t --compress-recursion=0 --profile-summary-cold-count=0
 ; RUN: FileCheck %s --input-file %t -check-prefix=CHECK-UNCOMPRESS
 ; RUN: llvm-profgen --format=text --perfscript=%S/Inputs/recursion-compression-pseudoprobe.perfscript --binary=%S/Inputs/recursion-compression-pseudoprobe.perfbin --output=%t --show-unwinder-output --profile-summary-cold-count=0 | FileCheck %s --check-prefix=CHECK-UNWINDER
 ; RUN: FileCheck %s --input-file %t
 ; RUN: llvm-profgen --format=text --perfscript=%S/Inputs/recursion-compression-pseudoprobe-nommap.perfscript --binary=%S/Inputs/recursion-compression-pseudoprobe.perfbin --output=%t --show-unwinder-output --profile-summary-cold-count=0 | FileCheck %s --check-prefix=CHECK-UNWINDER
 ; RUN: FileCheck %s --input-file %t
-
+; RUN: llvm-profgen --format=text --perfscript=%S/Inputs/recursion-compression-pseudoprobe.perfscript --binary=%S/Inputs/recursion-compression-pseudoprobe.perfbin --output=%t --compress-recursion=0 --profile-summary-cold-count=0 --csprof-max-context-depth=0
+; RUN: FileCheck %s --input-file %t -check-prefix=CHECK-MAX-CTX-DEPTH
 
 ; CHECK-UNCOMPRESS: [main:2 @ foo:5 @ fa:8 @ fa:7 @ fb:5 @ fb:5 @ fb:5 @ fb:5 @ fb:5 @ fb:5 @ fb:5 @ fb:5 @ fb:6 @ fa:8 @ fa:7 @ fb:6 @ fa]:4:1
 ; CHECK-UNCOMPRESS:  1: 1
 ; CHECK-UNCOMPRESS:  3: 1
 ; CHECK-UNCOMPRESS:  4: 1
 ; CHECK-UNCOMPRESS:  7: 1 fb:1
 ; CHECK-UNCOMPRESS:  !CFGChecksum: 563070469352221
 ; CHECK-UNCOMPRESS: [main:2 @ foo:5 @ fa:8 @ fa:7 @ fb:5 @ fb:5 @ fb:5 @ fb:5 @ fb:5 @ fb:5 @ fb:5 @ fb:5 @ fb:6 @ fa:8 @ fa]:4:1

 ; CHECK-UNCOMPRESS:  1: 1
 ; CHECK-UNCOMPRESS:  3: 1
 ; CHECK-UNCOMPRESS:  !CFGChecksum: 563070469352221
 ; CHECK-UNCOMPRESS: [main:2 @ foo:5 @ fa:8 @ fa:7 @ fb:5 @ fb:5 @ fb:5 @ fb:5 @ fb]:1:0
 ; CHECK-UNCOMPRESS:  5: 1 fb:1
 ; CHECK-UNCOMPRESS:  !CFGChecksum: 563022570642068
 
 
+; CHECK-MAX-CTX-DEPTH: [fb]:19:6
+; CHECK-MAX-CTX-DEPTH:  1: 6
+; CHECK-MAX-CTX-DEPTH:  2: 3
+; CHECK-MAX-CTX-DEPTH:  3: 3
+; CHECK-MAX-CTX-DEPTH:  4: 0
+; CHECK-MAX-CTX-DEPTH:  5: 4 fb:4
+; CHECK-MAX-CTX-DEPTH:  6: 3 fa:3
+; CHECK-MAX-CTX-DEPTH:  !CFGChecksum: 563022570642068
+; CHECK-MAX-CTX-DEPTH: [fa]:14:4
+; CHECK-MAX-CTX-DEPTH:  1: 4
+; CHECK-MAX-CTX-DEPTH:  3: 4
+; CHECK-MAX-CTX-DEPTH:  4: 2
+; CHECK-MAX-CTX-DEPTH:  5: 1
+; CHECK-MAX-CTX-DEPTH:  6: 0
+; CHECK-MAX-CTX-DEPTH:  7: 2 fb:2
+; CHECK-MAX-CTX-DEPTH:  8: 1 fa:1
+; CHECK-MAX-CTX-DEPTH:  !CFGChecksum: 563070469352221
+
+
 ; CHECK: [main:2 @ foo:5 @ fa:8 @ fa:7 @ fb:5 @ fb]:13:4
 ; CHECK:  1: 4
 ; CHECK:  2: 3
 ; CHECK:  3: 1
 ; CEHCK:  5: 4 fb:4
 ; CHECK:  6: 1 fa:1
 ; CHECK !CFGChecksum: 563022570642068
 ; CHECK: [main:2 @ foo:5 @ fa:8 @ fa:7 @ fb:5 @ fb:6 @ fa:8 @ fa:7 @ fb:6 @ fa]:6:2

llvm/tools/llvm-profgen/PerfReader.cpp

 
 static cl::opt<bool> ShowUnwinderOutput("show-unwinder-output",
                                         cl::ReallyHidden, cl::init(false),
                                         cl::ZeroOrMore,
                                         cl::desc("Print unwinder output"));
 
 extern cl::opt<bool> ShowDisassemblyOnly;
 extern cl::opt<bool> ShowSourceLocations;
 
 namespace llvm {
 namespace sampleprof {
 
 void VirtualUnwinder::unwindCall(UnwindState &State) {
   // The 2nd frame after leaf could be missing if stack sample is
   // taken when IP is within prolog/epilog, as frame chain isn't
   // setup yet. Fill in the missing frame in that case.
   // TODO: Currently we just assume all the addr that can't match the

 std::shared_ptr<ProbeBasedCtxKey> ProbeStack::getContextKey() {
   std::shared_ptr<ProbeBasedCtxKey> ProbeBasedKey =
       std::make_shared<ProbeBasedCtxKey>();
   for (auto CallProbe : Stack) {
     ProbeBasedKey->Probes.emplace_back(CallProbe);
   }
   CSProfileGenerator::compressRecursionContext<const MCDecodedPseudoProbe *>(
       ProbeBasedKey->Probes);
+  CSProfileGenerator::trimContext<const MCDecodedPseudoProbe *>(
+      ProbeBasedKey->Probes);
+
   ProbeBasedKey->genHashCode();
   return ProbeBasedKey;
 }
 
 template <typename T>
 void VirtualUnwinder::collectSamplesFromFrame(UnwindState::ProfiledFrame *Cur,
                                               T &Stack) {
   if (Cur->RangeSamples.empty() && Cur->BranchSamples.empty())

llvm/tools/llvm-profgen/ProfileGenerator.h

 
 public:
   CSProfileGenerator(const BinarySampleCounterMap &Counters)
       : BinarySampleCounters(Counters){};
 
 public:
   void generateProfile() override;
 
+  // Trim the context stack at a given depth.
+  template <typename T>
+  static void trimContext(SmallVectorImpl<T> &S, int Depth = MaxContextDepth) {
+    if (Depth < 0 || static_cast<size_t>(Depth) >= S.size())
+      return;
+    std::copy(S.begin() + S.size() - static_cast<size_t>(Depth), S.end(),
+              S.begin());
+    S.resize(Depth);
+  }
+
   // Remove adjacent repeated context sequences up to a given sequence length,
   // -1 means no size limit. Note that repeated sequences are identified based
   // on the exact call site, this is finer granularity than function recursion.
   template <typename T>
   static void compressRecursionContext(SmallVectorImpl<T> &Context,
                                        int32_t CSize = MaxCompressionSize) {
     uint32_t I = 1;
     uint32_t HS = static_cast<uint32_t>(Context.size() / 2);
void populateFunctionBoundarySamples(StringRef ContextId,
                                        const BranchSample &BranchCounters,
                                        ProfiledBinary *Binary);
   void populateInferredFunctionSamples();
 
 public:
   // Deduplicate adjacent repeated context sequences up to a given sequence
   // length. -1 means no size limit.
   static int32_t MaxCompressionSize;
+  static int MaxContextDepth;
 };
 
 using ProbeCounterMap =
     std::unordered_map<const MCDecodedPseudoProbe *, uint64_t>;
 
 class PseudoProbeCSProfileGenerator : public CSProfileGenerator {
 
 public:

llvm/tools/llvm-profgen/ProfileGenerator.cpp

cl::desc("If the total count of context profile is smaller than "
              "the threshold, it will be merged into context-less base "
              "profile."));
 
 static cl::opt<bool> CSProfTrimColdContext(
     "csprof-trim-cold-context", cl::init(true), cl::ZeroOrMore,
     cl::desc("If the total count of the profile after all merge is done "
              "is still smaller than threshold, it will be trimmed."));
 
-static cl::opt<uint32_t> CSProfColdContextFrameDepth(
-    "csprof-frame-depth-for-cold-context", cl::init(1), cl::ZeroOrMore,
-    cl::desc("Keep the last K frames while merging cold profile. 1 means the "
+static cl::opt<uint32_t> CSProfMaxColdContextDepth(
+    "csprof-max-cold-context-depth", cl::init(1), cl::ZeroOrMore,
+    cl::desc("Keep the last K contexts while merging cold profile. 1 means the "
              "context-less base profile"));
 
+static cl::opt<int, true> CSProfMaxContextDepth(
+    "csprof-max-context-depth", cl::ZeroOrMore,
+    cl::desc("Keep the last K contexts while merging profile. -1 means no "
+             "depth limit."),
+    cl::location(llvm::sampleprof::CSProfileGenerator::MaxContextDepth));
+
 static cl::opt<bool> EnableCSPreInliner(
     "csspgo-preinliner", cl::Hidden, cl::init(false),
     cl::desc("Run a global pre-inliner to merge context profile based on "
              "estimated global top-down inline decisions"));
 
 extern cl::opt<int> ProfileSummaryCutoffCold;
 
 using namespace llvm;
 using namespace sampleprof;
 
 namespace llvm {
 namespace sampleprof {
 
 // Initialize the MaxCompressionSize to -1 which means no size limit
 int32_t CSProfileGenerator::MaxCompressionSize = -1;
 
+int CSProfileGenerator::MaxContextDepth = -1;
+
 static bool
 usePseudoProbes(const BinarySampleCounterMap &BinarySampleCounters) {
   return BinarySampleCounters.size() &&
          BinarySampleCounters.begin()->first->usePseudoProbes();
 }
 
 std::unique_ptr<ProfileGenerator>
 ProfileGenerator::create(const BinarySampleCounterMap &BinarySampleCounters,
void CSProfileGenerator::postProcessProfiles() {
   // inline decisions.
   if (EnableCSPreInliner)
     CSPreInliner(ProfileMap, HotCountThreshold, ColdCountThreshold).run();
 
   // Trim and merge cold context profile using cold threshold above;
   SampleContextTrimmer(ProfileMap)
       .trimAndMergeColdContextProfiles(
           ColdCountThreshold, CSProfTrimColdContext, CSProfMergeColdContext,
-          CSProfColdContextFrameDepth);
+          CSProfMaxColdContextDepth);
 }
 
 void CSProfileGenerator::computeSummaryAndThreshold() {
   // Update the default value of cold cutoff for llvm-profgen.
   // Do it here because we don't want to change the global default,
   // which would lead CS profile size too large.
   if (!ProfileSummaryCutoffCold.getNumOccurrences())
     ProfileSummaryCutoffCold = 999000;

 FunctionSamples &PseudoProbeCSProfileGenerator::getFunctionProfileForLeafProbe(
     SmallVectorImpl<std::string> &ContextStrStack,
     const MCPseudoProbeFuncDesc *LeafFuncDesc, bool WasLeafInlined) {
   assert(ContextStrStack.size() && "Profile context must have the leaf frame");
   // Compress the context string except for the leaf frame
   std::string LeafFrame = ContextStrStack.back();
   ContextStrStack.pop_back();
   CSProfileGenerator::compressRecursionContext(ContextStrStack);
+  CSProfileGenerator::trimContext(ContextStrStack);
 
   std::ostringstream OContextStr;
   for (uint32_t I = 0; I < ContextStrStack.size(); I++) {
     if (OContextStr.str().size())
       OContextStr << " @ ";
     OContextStr << ContextStrStack[I];
   }
   // For leaf inlined context with the top frame, we should strip off the top

llvm/tools/llvm-profgen/ProfiledBinary.cpp

for (auto Address : Stack) {
     }
   }
 
   assert(ContextVec.size() && "Context length should be at least 1");
   // Compress the context string except for the leaf frame
   std::string LeafFrame = ContextVec.back();
   ContextVec.pop_back();
   CSProfileGenerator::compressRecursionContext<std::string>(ContextVec);
+  CSProfileGenerator::trimContext<std::string>(ContextVec);
 
   std::ostringstream OContextStr;
   for (uint32_t I = 0; I < (uint32_t)ContextVec.size(); I++) {
     if (OContextStr.str().size()) {
       OContextStr << " @ ";
     }
     OContextStr << ContextVec[I];
   }